Avoid Overthinking in Self-Supervised Models for Speech Recognition

TL;DR

This paper investigates overthinking in self-supervised speech recognition models, demonstrating the issue and proposing two novel early exit strategies to improve inference efficiency and accuracy, especially on out-of-distribution data.

Contribution

It introduces the first analysis of overthinking in SSL-based ASR and proposes two new early exit strategies tailored for speech recognition tasks.

Findings

SSL models exhibit overthinking in ASR.

Proposed strategies outperform previous early exit methods.

Optimal performance-speed trade-off bounds are computed.

Abstract

Self-supervised learning (SSL) models reshaped our approach to speech, language and vision. However their huge size and the opaque relations between their layers and tasks result in slow inference and network overthinking, where predictions made from the last layer of large models is worse than those made from intermediate layers. Early exit (EE) strategies can solve both issues by dynamically reducing computations at inference time for certain samples. Although popular for classification tasks in vision and language, EE has seen less use for sequence-to-sequence speech recognition (ASR) tasks where outputs from early layers are often degenerate. This challenge is further compounded when speech SSL models are applied on out-of-distribution (OOD) data. This paper first shows that SSL models do overthinking in ASR. We then motivate further research in EE by computing an optimal bound for performance versus speed trade-offs. To approach this bound we propose two new strategies for ASR: (1) we adapt the recently proposed patience strategy to ASR; and (2) we design a new EE strategy specific to ASR that performs better than all strategies previously introduced.